Rethinking Driving Topology Reasoning: Plug-and-Play Discrete Graph Refinement

ICLR 2026 Conference Submission14075 Authors

Keywords: Online Map Generation, Autonomous Driving, Topology Graph Refinement, Self-Supervised Learning, Heterogeneous GNN

Abstract: In autonomous driving, topology reasoning aims to recover the structured connectivity of road networks by detecting map elements and predicting their relations, providing machine-readable maps for safe and efficient operation. Surprisingly, current topology reasoning tasks do not address how to produce better discrete graphs, even though downstream modules such as planning and control rely on them. Existing methods predict continuous edge scores and then apply simple thresholding to obtain discrete graphs, but this step is neither optimized during training nor evaluated in benchmarks. As a result, it remains unclear whether their predicted continuous graphs are truly effective for downstream tasks. To bridge this gap, we propose **TopoRefine**, a universal and plug-and-play topology graph refinement module that refines continuous graphs predicted by any topology reasoning model into higher-quality discrete graphs. Specifically, it refines connectivity by learning structural patterns via a lightweight GNN-based refinement module trained in a self-supervised way. This refinement module calibrates predictions so that thresholding yields more reliable discrete structures. In addition, we are the first to introduce a discrete graph evaluation metric in this setting, the Topology Jaccard Score, tailored to directly assess the quality of discrete driving topology graph. Experiments on multiple baselines demonstrate that TopoRefine improves both continuous and discrete graph quality, making it the first framework to explicitly focus on improving discrete graph reliability in topology reasoning.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 14075